MULTIMODAL DATA FUSION APPROACH FOR ACCURATE KIDNEY CANCER PROGNOSIS

Abstract

Kidney cancer, especially renal cell carcinoma (RCC) is a life threatening disease, heterogeneous in tumor appearance, complicated in biological behavior, and diagnosed at an advanced stage. Prognosis can only be accurately done by combining the radiological pattern with clinical and laboratory biomarkers. This thesis presents Multimodal Data Fusion Approach to Accurate Prognosis of kidney cancer, which is a proposed method of integrating medical imaging and structured clinical data to achieve more accurate prediction and clinical decision making. The computed tomography (CT) kidney scans in Kaggle kidney cancer were used, as well as clinical information (age, tumor stage, biomarkers, and laboratory indicators). Convolutional Neural Network (CNN) was utilized to derive deep spatial information on kidney CT images, including tumor shape, texture, and changes in intensity. Parallel to them, the Random Forest (RF) and the Logistic Regression (LR) models were used on structured clinical data that had been imputed, coded, and scaled. The multimodal fusion approach was used to combine CNN-learned imaging embeddings with clinical model predictions with a late fusion approach. Accuracy, precision, recall, F1-score, confusion matrices, and training-validation loss curves were used as the performance measures of the models. The CNN recorded the best classification score of 99.4, which is good at acquiring hierarchical features. Random Forest was found to be effective in modelling nonlinear clinical interactions with 95.2, whereas Logistic Regression was found to be better, with 96.6, and provided interpretable risk estimation. It proves that integration of imaging intelligence with clinical background are much better to enhance the accuracy, robustness and clinical relevance in pronging kidney cancer.